The occurrence of sleep is widespread over the large majority of animal species, suggesting a specific evolutionary strategy. The activity displayed in such a state is known to be beneficial for cognitive functions, stabilizing memories and improving the performances in several tasks. Despite this, a theoretical and computational approach to achieve the understanding of such mechanism is still lacking. In this paper we show the effect of sleep-like activity on a simplified thalamo-cortical model which is trained to encode and retrieve images of handwritten digits. We show that, if spike-timing-dependent-plasticity (STDP) is active during sleep, the connections among groups of neurons associated to instances of the same class (digit) are enhanced, and that the internal representation is hierarchical and orthogonalized. Such effect might be beneficial to the network to obtain better performances in retrieval and classification tasks and to create hierarchies of categories in integrated representations. The model leverages on the coincidence of top-down contextual information with bottom-up sensory flow during the training phase and on the integration of top-down predictions and bottom-up pathways during deep-sleep-like slow oscillations.